Where do agent costs hide?
Agent costs hide in the time of the people around the agent. An agent that acts without a gate has to be watched; one that produces work without evidence has to be checked; one with broad credentials has to be trusted; one with no record of its decisions has to be re-examined when something goes wrong. None of that appears on the invoice.
The Digital FTE model on ftes.ai is designed to move those costs into the open, where they can be measured and reduced, rather than to pretend they do not exist.
What is the supervision cost?
The supervision cost is the manager's time spent reading everything an agent does because there is no structured point at which to review it. Without an approval gate, supervision means reading outputs after the fact, which is slower and catches errors later.
With an approval gate, supervision is the review of proposals before they act, in a queue that puts what is waiting first. It is still time, but it is measurable, it falls as the approval rate rises, and it ends for classes of action that earn autonomy.
What is the rework cost?
The rework cost is the time to undo an error that was not caught: a published claim that was wrong, a ledger entry that should not have posted, a candidate email that should not have gone. Rework is expensive because it includes the damage as well as the correction.
The fact gate and the evidence-on-every-proposal rule reduce rework by catching errors before they act. An unsourced claim does not pass the gate; a proposed entry shows its evidence; a candidate email waits for approval. The cost moves from rework to review, which is cheaper.
- Supervision: reading after the fact versus reviewing proposals before
- Rework: undoing errors versus catching them at the gate
- Credentials: broad standing access versus scoped, revocable, logged
- Drift: noticed when something breaks versus visible in the approval record
What is the credential cost?
The credential cost is the risk carried by agents holding broad access that nobody tracks. An agent with a shared admin login to the CRM, the ledger, and the email system is an incident waiting for a cause, and the cost of that incident does not appear until it happens.
Scoped credentials, granted per tool and per property, revocable from one place, and logged in the audit trail, turn that hidden risk into a visible control. The credential management guide on ftes.ai describes the model.
What is the drift cost?
The drift cost is what happens when an agent's behaviour changes over time and nobody notices until an outcome is wrong. Models update, sources change, and an agent with no record of what it proposed and what was approved cannot be compared against its past self.
The approval record is the instrument for noticing drift. A category whose approval rate falls, or whose send-back reasons change, shows drift before it becomes an incident, and the manager can step the category back on the autonomy ladder while the cause is found.
How do you measure the hidden costs before and after?
Measure the hidden costs by logging the human time they consume. Before an AI employee arrives, or with an ungoverned agent already running, record for two weeks the hours spent reading its outputs after the fact, the hours spent undoing anything it got wrong, the number of credentials it holds and who last reviewed them, and any change in its behaviour that nobody noticed until an outcome was wrong. Most teams have never added these up, and the total is usually the surprise.
With a Digital FTE on ftes.ai the same four items become visible numbers rather than hidden ones: approval minutes per proposal in the audit trail, send-backs and their reasons in place of rework, the connections listed per role in Setup with their scope and last change, and the approval rate per category as the drift detector. Compare the two records after two cycles. The point is not that the governed model has no cost; it is that its cost is measured, falls as the record grows, and never lands silently on the people around the agent.